System and Method for Recurrent Neural Networks for Forecasting of Consumer Goods' Sales and Inventory

1. An integrated forecasting system with computerized components for forecasting a future performance value based on time series data corresponding to at least one past performance, the system comprising: a. a training input routine executing at a microprocessor coupled to a non-transitory media storing instructions for executing the training input routine, the training input routine configured to: i. receive at least one data input message comprising an array of time series data comprising a plurality of data points, wherein each of the one or more data points comprises a single record of the time series data, further wherein each of the one or more data points comprises a date field, a value field, and a meta field; ii. for each of the date field, value field, and meta field, determine whether the particular field comprises a known value or an unknown value; iii. for each data point wherein it is determined that at least one of the date field, value field, and meta field for the particular data point comprises an unknown value, mark the particular data point as a context data point; and iv. for each data point wherein it is determined that each of the data field, value field, and meta field comprises a known value, mark the particular data point as a training data point; b. a training scheduler executing at the microprocessor coupled to the non-transitory media storing instructions for executing the training scheduler, the training scheduler configured to: i. receive from the training input routine the times series data comprising the plurality of data points; ii. calculate a frequency associated with the time series data based on the known data field of each of the data points; iii. normalize the known value fields of each of the data points, wherein normalizing the known value fields comprises determining a constant scale that when multiplied by an absolute value of the known value fields of each of the data points results in a normalized value for each data point less than or equal to one; iv. extract metadata from the meta field of each data point and reproduce the metadata from the meta field of each data point into the meta field of each other data point such that the meta field of all data points comprise all of the metadata from each data point; and v. prepare a list of training jobs, the list comprising a number of training jobs, wherein the list of training jobs is assigned an ensemble ID, further wherein each of the training jobs is assigned a training job ID c. a job queue in communication with the training scheduler, the job queue configured to receive from the training scheduler the list of training jobs in a particular order and store the list of training jobs in the particular order received from the training scheduler for retrieval; d. a training worker executing at the microprocessor coupled to the non-transitory media storing instructions for executing the training worker, the training worker configured to fetch each of the training jobs in the list of training jobs from the job queue in the particular order received from the training scheduler, and for each of the training jobs: i. assign each of the data points as one of past data and future data, wherein past data corresponds to training data value fields multiplied by the constant scale and meta values associated with the next data point, and wherein future data corresponds to context data meta fields shifted such that the ith entry is the meta value for the (i+1)th forecasted point; ii. train one or more neural networks using the past data; and iii. generate a sequence of predictions using the one or more trained neural networks, past data, and future data, wherein each of the predictions comprises a date, a value, and an ensemble ID, wherein the ensemble ID of the prediction corresponds to the ensemble ID assigned to the training jobs by the training worker; e. a predictions database in communication with the training worker and configured to receive from the training worker the sequence of predictions and store the predictions for retrieval; f. a predictor input routine executing at the microprocessor coupled to the non-transitory media storing instructions for executing the predictor input routine, the predictor input routine configured to receive a predictor input message and read from the predictor input message an input ensemble ID, a start date, and an end date; g. a predictor executing at the microprocessor coupled to the non-transitory media storing instructions for executing the predictor, the predictor configured to: i. receive from the predictor input routine the input ensemble ID, start date, and end date; ii. fetch from the predictions database all of the stored predictions having an ensemble ID matching the input ensemble ID; iii. filter the fetched predictions such that only the predictions having a date within a time range defined by the input start date and input end date are obtained; iv. aggregate the filtered predictions; v. determine a mean predicted value and a standard deviation value for the aggregate predictions wherein the mean predicted value is output as the forecasted future performance value and the standard deviation value is output as a confidence metric for the forecasted future performance value.

2. The system of claim 1 , wherein the training input routine is further configured to receive an ensemble size input message and read from the ensemble size input message an ensemble size providing the number of training jobs to be prepared by the training scheduler.

3. The system of claim 1 , wherein the training input routine is further configured to receive one or more options input messages and to read from each of the one or more options input messages one or more option inputs providing a number of hyperparameters to be used by the training worker to train the one or more neural networks.

4. The system of claim 1 , wherein the date field of each data point includes a timestamp corresponding to a date when the single record associated with the particular data point occurred.

5. The system of claim 1 , wherein the value field of a number of the data points includes a numeric value corresponding to an observed value of a particular characteristic associated with the particular data point.

6. The system of claim 1 , wherein the meta field of each data point includes information denoting environmental factors associated with the record corresponding to the particular data point.

7. A computerized method for forecasting a future performance value based on time series data corresponding to at least one past performance, the method comprising the steps of: a. receiving one or more training input messages at a microprocessor, wherein each of the training input messages comprises an array of time series data points, wherein each of the data points comprises a date field, a value field, and a meta field, further wherein each of the date field, value field, and meta field is one of a known value or an unknown value; b. for each of the date field, value field, and meta field of each data point, determining whether the particular field comprises a known value or an unknown value; c. for each data point wherein it is determined that at least one of the date field, value field, and meta field for the particular data point comprises an unknown value, mark the particular data point as a context data point; and d. for each data point wherein it is determined that each of the data field, value field, and meta field comprises a known value, mark the particular data point as a training data point; e. preparing a list of training jobs, the list comprising a number of individual training jobs, wherein the list of training jobs is assigned an ensemble ID, further wherein each of the individual training jobs is assigned a training job ID; f. processing the list of training jobs starting with the first training job in the list and continuing through the list of training jobs, the processing of each training job comprising training one or more neural networks using the training data and context data associated with the particular training job to determine a predicted value associated with the particular training job; g. storing the predicted value associated with each of the particular training jobs in a predictions database, wherein the predicted value of each of the training jobs is paired with a date associated with the particular training job and the ensemble ID and job ID associated with the particular training job; h. receiving a prediction input message comprising a prediction ensemble ID, a i. fetching from the predictions database all predicted values having an ensemble ID matching the input prediction ensemble ID; j. aggregating all of the fetched predicted values; k. calculating a mean predicted value from the aggregated predicted values; and l. calculating a standard deviation from the aggregated predicted values; wherein the calculated mean predicted value is output as the forecasted future performance value and wherein the calculated standard deviation is output as a confidence metric associated with the forecasted future performance value.